🛰️ Daily AI Frontier
‹ back to 2026-08-28

UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City

Research LLM Agents

Ranking

Overall 87
Content 95
Popularity 69

Observed public metrics from 1 member.

Representative image for UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City

Merged summary

TL;DR - UrbanGround is a closed-loop benchmark for testing whether multimodal LLM agents can turn street-level perception into reliable navigation in a physically constrained 3D replica of Hong Kong. Current agents handle basic visual recognition and short-range spatial reasoning, but struggle to sustain and correct goal-directed behavior over longer routes.

  • Built from territory-wide 3D geospatial data, the sandbox supports first-person exploration and interactive-map navigation.
  • Evaluates active spatial grounding, navigation to increasingly distant or ambiguous destinations, and robustness to route changes and pedestrian motion.
  • Orientation and pedestrian-aware movement remain unreliable despite useful local perception capabilities.
  • During extended exploration, errors accumulate because agents fail to compose local skills into sustained behavior or recover effectively.

Sources (1)

UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City

arXiv cs.CV Tianjie Ju, Zheng Wu, Yueqing Sun, Yuhan Cui, Bobo Li, Shengqiong Wu, Pengzhou Cheng, Haodong Zhao, Zongru Wu, Xinbei Ma, Doris Zhang, Kunling Li, Mong-Li Lee, Wynne Hsu, Hao Fei, Qi Gu, Gongshen Liu, Zhuosheng Zhang 2026-08-27 arXiv:2608.27456
Public signals Hugging Face upvotes 86
Providers: Hugging Face · Upvotes 86 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-25 14:27:05.639123 UTC

TL;DR - UrbanGround is a closed-loop benchmark for testing whether multimodal LLM agents can turn street-level perception into reliable navigation in a physically constrained 3D replica of Hong Kong. Current agents handle basic visual recognition and short-range spatial reasoning, but struggle to sustain and correct goal-directed behavior over longer routes.

  • Built from territory-wide 3D geospatial data, the sandbox supports first-person exploration and interactive-map navigation.
  • Evaluates active spatial grounding, navigation to increasingly distant or ambiguous destinations, and robustness to route changes and pedestrian motion.
  • Orientation and pedestrian-aware movement remain unreliable despite useful local perception capabilities.
  • During extended exploration, errors accumulate because agents fail to compose local skills into sustained behavior or recover effectively.
item →